GNNFairViz: Visual Analysis for Graph Neural Network Fairness.

Area of Science:

• Artificial Intelligence
• Data Science
• Human-Computer Interaction

Background:

• Graph Neural Networks (GNNs) show great potential but raise fairness concerns, especially in human-centric applications, risking discrimination.
• Existing visual analytics for machine learning (ML) fairness often overlook the unique challenges presented by GNNs.
• Attribute and structural biases in GNNs can lead to significant model bias, necessitating specialized analysis tools.

Purpose of the Study:

• To propose a novel visual analytics framework for analyzing and mitigating fairness issues in Graph Neural Networks (GNNs).
• To provide insights into how attribute and structural biases contribute to model bias in GNNs.
• To develop an operational tool, GNNFairViz, for GNN developers to proactively assess and address fairness.

Main Methods:

• Developed a model-agnostic visual analytics framework for GNN fairness analysis, supporting multiple sensitive attributes.
• Created GNNFairViz, an interactive visual analysis tool integrated into the GNN development workflow.
• Utilized an extended suite of fairness metrics for comprehensive bias inspection and diagnostics.

Main Results:

• GNNFairViz enables developers to effectively analyze GNN bias, select nodes, and perform fairness inspections.
• Evaluation through usage scenarios and expert interviews confirmed the framework's effectiveness and usability.
• Identified the 'Overwhelming Effect' in unbalanced datasets and emphasized the role of GNN architecture in bias mitigation.

Conclusions:

• The proposed visual analytics framework and GNNFairViz tool significantly enhance the ability to analyze and address fairness in GNNs.
• The findings offer practical guidance for developing fairer GNN models in real-world applications.
• Further research into GNN fairness should consider dataset imbalance and architectural choices for effective bias mitigation.